Meta unveils two frontier AI models, Muse Wonder and Muse Wonder Full, to challenge OpenAI and Anthropic

A post reports that Meta has introduced two frontier AI models, Muse Wonder and Muse Wonder Full, in a move aimed at competing with OpenAI and Anthropic. The post, by an account named Bearlovesbull, describes the release as one of the biggest tests of the billions Meta CEO Mark Zuckerberg has invested in AI. The models were trained using Meta's gigawatt-scale Prometheus data center, and Meta claims they can match or beat the latest models from OpenAI and Anthropic across several benchmarks, according to the post. The post also highlights two features it says are more interesting: Active Memory, which helps agents remember information about users across conversations, and Active Focus, which helps them stay on a task, try different approaches, and adjust when something goes wrong. The post suggests these features could enable agents working across WhatsApp, Instagram, and Facebook, allowing them to remember user preferences and continue working through multiple steps without repeated explanations. The post notes that Meta already has distribution and does not need to convince billions of people to install another application, but the question was whether Zuckerberg could build models competitive enough to take advantage of that distribution. The post says today's announcement explains Meta is making serious progress, although the performance claims still need independent testing. According to the post, Muse Wonder is available through Meta's API and Muse Code, while Wonder Full is available to Muse Code subscribers, and Wonder will also power future Muse experiences. The post adds that if the models perform as advertised, Meta could gradually reduce its dependence on outside AI providers while controlling more of the intelligence behind its own products. However, the post notes that Meta has not committed to releasing the Wonder models as open weights, which it finds interesting given the company's history of promoting Llama as an open model alternative. The post concludes that Meta will try to keep its most capable agent models close to its own products, especially where memory, personalization, and user relationships give it a competitive advantage, and that Meta now wants both strong models and distribution, as OpenAI and Anthropic have strong models and Google has its own models and distribution. The post also mentions the need for independent benchmarks, actual Muse retention, and whether people start trusting these agents with more complicated tasks, as getting users to try an AI agent is relatively easy when you own WhatsApp and Instagram, but getting them to depend on it every day is where the real business opportunity begins.